arXiv subjects

Xurui Song

Publications and source records attributed to Xurui Song.

9 recordsLinked to original sources

arXiv2026

Visual Attention Faithfulness in Vision-Language Models is Heterogeneous

Whether attention weights faithfully reflect model reasoning has been actively debated in NLP, yet this question remains largely unexplored for the visual modality in Vision-Language Models (VLMs). We address this gap through causal perturbation analysis on current VLMs, evaluating both the comprehensiveness and sufficiency gap of attention-ranked visual tokens. Our analysis reveals that visual attention faithfulness is heterogeneous, manifesting in three distinct processing modes: Faithful-Sufficient, where top-$k$ attention tokens are both necessary and sufficient for prediction; Faithful-Distributed, where they are necessary but broader visual context remains required; and Non-Focal, where no localized attention region is individually necessary while visual information remains an essential trigger for prediction. Furthermore, human-annotated ground-truth regions satisfy comprehensiveness in only $\sim 60$% of cases compared with model attention rankings, revealing systematic divergence between model visual reliance and human intuition. We demonstrate these patterns across both general VQA on VQAv2 and document tasks on VRDU and ChartQA, showing that visual attention faithfulness varies systematically with processing demands and model architectures rather than being uniformly faithful or unfaithful.

cs.CV↗

arXiv2026

TrojanPraise: Jailbreak LLMs via Benign Fine-Tuning

The demand of customized large language models (LLMs) has led to commercial LLMs offering black-box fine-tuning APIs, yet this convenience introduces a critical security loophole: attackers could jailbreak the LLMs by fine-tuning them with malicious data. Though this security issue has recently been exposed, the feasibility of such attacks is questionable as malicious training dataset is believed to be detectable by moderation models such as Llama-Guard-3. In this paper, we propose TrojanPraise, a novel finetuning-based attack exploiting benign and thus filter-approved data. Basically, TrojanPraise fine-tunes the model to associate a crafted word (e.g., "bruaf") with harmless connotations, then uses this word to praise harmful concepts, subtly shifting the LLM from refusal to compliance. To explain the attack, we decouple the LLM's internal representation of a query into two dimensions of knowledge and attitude. We demonstrate that successful jailbreak requires shifting the attitude while avoiding knowledge shift, a distortion in the model's understanding of the concept. To validate this attack, we conduct experiments on five opensource LLMs and two commercial LLMs under strict black-box settings. Results show that TrojanPraise achieves a maximum attack success rate of 95.88% while evading moderation.

cs.CR↗

arXiv2025

Cheating Stereo Matching in Full-scale: Physical Adversarial Attack against Binocular Depth Estimation in Autonomous Driving

Though deep neural models adopted to realize the perception of autonomous driving have proven vulnerable to adversarial examples, known attacks often leverage 2D patches and target mostly monocular perception. Therefore, the effectiveness of Physical Adversarial Examples (PAEs) on stereo-based binocular depth estimation remains largely unexplored. To this end, we propose the first texture-enabled physical adversarial attack against stereo matching models in the context of autonomous driving. Our method employs a 3D PAE with global camouflage texture rather than a local 2D patch-based one, ensuring both visual consistency and attack effectiveness across different viewpoints of stereo cameras. To cope with the disparity effect of these cameras, we also propose a new 3D stereo matching rendering module that allows the PAE to be aligned with real-world positions and headings in binocular vision. We further propose a novel merging attack that seamlessly blends the target into the environment through fine-grained PAE optimization. It has significantly enhanced stealth and lethality upon existing hiding attacks that fail to get seamlessly merged into the background. Extensive evaluations show that our PAEs can successfully fool the stereo models into producing erroneous depth information.

cs.CV↗

arXiv2025

More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models

Vision-Language Model (VLM) driving agents promise explainable end-to-end autonomy by first producing natural-language reasoning and then predicting trajectory planning. However, whether planning is causally driven by this reasoning remains a critical but unverified assumption. To investigate this, we build DriveMind, a large-scale driving Visual Question Answering (VQA) corpus with plan-aligned Chain-of-Thought (CoT), automatically generated from nuPlan. Our data generation process converts sensors and annotations into structured inputs and, crucially, separates priors from to-be-reasoned signals, enabling clean information ablations. Using DriveMind, we train representative VLM agents with Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) and evaluate them with nuPlan's metrics. Our results, unfortunately, indicate a consistent causal disconnect in reasoning-planning: removing ego/navigation priors causes large drops in planning scores, whereas removing CoT produces only minor changes. Attention analysis further shows that planning primarily focuses on priors rather than the CoT. Based on this evidence, we propose the Reasoning-Planning Decoupling Hypothesis, positing that the training-yielded reasoning is an ancillary byproduct rather than a causal mediator. To enable efficient diagnosis, we also introduce a novel, training-free probe that measures an agent's reliance on priors by evaluating its planning robustness against minor input perturbations. In summary, we provide the community with a new dataset and a diagnostic tool to evaluate the causal fidelity of future models.

cs.AI↗

arXiv2025

Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs

Despite substantial efforts in safety alignment, recent research indicates that Large Language Models (LLMs) remain highly susceptible to jailbreak attacks. Among these attacks, finetuning-based ones that compromise LLMs' safety alignment via fine-tuning stand out due to its stable jailbreak performance. In particular, a recent study indicates that fine-tuning with as few as 10 harmful question-answer (QA) pairs can lead to successful jailbreaking across various harmful questions. However, such malicious fine-tuning attacks are readily detectable and hence thwarted by moderation models. In this paper, we demonstrate that LLMs can be jailbroken by fine-tuning with only 10 benign QA pairs; our attack exploits the increased sensitivity of LLMs to fine-tuning data after being overfitted. Specifically, our fine-tuning process starts with overfitting an LLM via fine-tuning with benign QA pairs involving identical refusal answers. Further fine-tuning is then performed with standard benign answers, causing the overfitted LLM to forget the refusal attitude and thus provide compliant answers regardless of the harmfulness of a question. We implement our attack on the ten LLMs and compare it with five existing baselines. Experiments demonstrate that our method achieves significant advantages in both attack effectiveness and attack stealth. Our findings expose previously unreported security vulnerabilities in current LLMs and provide a new perspective on understanding how LLMs' security is compromised, even with benign fine-tuning. Our code is available at https://github.com/ZHIXINXIE/tenBenign.

cs.CR↗

arXiv2025

Where to Start Alignment? Diffusion Large Language Model May Demand a Distinct Position

Diffusion Large Language Models (dLLMs) have recently emerged as a competitive non-autoregressive paradigm due to their unique training and inference approach. However, there is currently a lack of safety study on this novel architecture. In this paper, we present the first analysis of dLLMs' safety performance and propose a novel safety alignment method tailored to their unique generation characteristics. Specifically, we identify a critical asymmetry between the defender and attacker in terms of security. For the defender, we reveal that the middle tokens of the response, rather than the initial ones, are more critical to the overall safety of dLLM outputs; this seems to suggest that aligning middle tokens can be more beneficial to the defender. The attacker, on the contrary, may have limited power to manipulate middle tokens, as we find dLLMs have a strong tendency towards a sequential generation order in practice, forcing the attack to meet this distribution and diverting it from influencing the critical middle tokens. Building on this asymmetry, we introduce Middle-tOken Safety Alignment (MOSA), a novel method that directly aligns the model's middle generation with safe refusals exploiting reinforcement learning. We implement MOSA and compare its security performance against eight attack methods on two benchmarks. We also test the utility of MOSA-aligned dLLM on coding, math, and general reasoning. The results strongly prove the superiority of MOSA.

cs.CR↗

arXiv2025

Corvid: Improving Multimodal Large Language Models Towards Chain-of-Thought Reasoning

Recent advancements in multimodal large language models (MLLMs) have demonstrated exceptional performance in multimodal perception and understanding. However, leading open-source MLLMs exhibit significant limitations in complex and structured reasoning, particularly in tasks requiring deep reasoning for decision-making and problem-solving. In this work, we present Corvid, an MLLM with enhanced chain-of-thought (CoT) reasoning capabilities. Architecturally, Corvid incorporates a hybrid vision encoder for informative visual representation and a meticulously designed connector (GateMixer) to facilitate cross-modal alignment. To enhance Corvid's CoT reasoning capabilities, we introduce MCoT-Instruct-287K, a high-quality multimodal CoT instruction-following dataset, refined and standardized from diverse public reasoning sources. Leveraging this dataset, we fine-tune Corvid with a two-stage CoT-formatted training approach to progressively enhance its step-by-step reasoning abilities. Furthermore, we propose an effective inference-time scaling strategy that enables Corvid to mitigate over-reasoning and under-reasoning through self-verification. Extensive experiments demonstrate that Corvid outperforms existing o1-like MLLMs and state-of-the-art MLLMs with similar parameter scales, with notable strengths in mathematical reasoning and science problem-solving. Project page: https://mm-vl.github.io/corvid.

cs.CV↗

arXiv2025

Dagger Behind Smile: Fool LLMs with a Happy Ending Story

The wide adoption of Large Language Models (LLMs) has attracted significant attention from $\textit{jailbreak}$ attacks, where adversarial prompts crafted through optimization or manual design exploit LLMs to generate malicious contents. However, optimization-based attacks have limited efficiency and transferability, while existing manual designs are either easily detectable or demand intricate interactions with LLMs. In this paper, we first point out a novel perspective for jailbreak attacks: LLMs are more responsive to $\textit{positive}$ prompts. Based on this, we deploy Happy Ending Attack (HEA) to wrap up a malicious request in a scenario template involving a positive prompt formed mainly via a $\textit{happy ending}$, it thus fools LLMs into jailbreaking either immediately or at a follow-up malicious request. This has made HEA both efficient and effective, as it requires only up to two turns to fully jailbreak LLMs. Extensive experiments show that our HEA can successfully jailbreak on state-of-the-art LLMs, including GPT-4o, Llama3-70b, Gemini-pro, and achieves 88.79% attack success rate on average. We also provide quantitative explanations for the success of HEA.

cs.CL↗

arXiv2023

Do Not DeepFake Me: Privacy-Preserving Neural 3D Head Reconstruction Without Sensitive Images

While 3D head reconstruction is widely used for modeling, existing neural reconstruction approaches rely on high-resolution multi-view images, posing notable privacy issues. Individuals are particularly sensitive to facial features, and facial image leakage can lead to many malicious activities, such as unauthorized tracking and deepfake. In contrast, geometric data is less susceptible to misuse due to its complex processing requirements, and absence of facial texture features. In this paper, we propose a novel two-stage 3D facial reconstruction method aimed at avoiding exposure to sensitive facial information while preserving detailed geometric accuracy. Our approach first uses non-sensitive rear-head images for initial geometry and then refines this geometry using processed privacy-removed gradient images. Extensive experiments show that the resulting geometry is comparable to methods using full images, while the process is resistant to DeepFake applications and facial recognition (FR) systems, thereby proving its effectiveness in privacy protection.

cs.CV↗